Economics Letters 20 (1986) 177-181 North-Holland

177

FINITE STATE MARKOV-CHAIN APPROXIMATIONS UNIVARIATE AND VECTOR AUTOREGRESSIONS George TAUCHEN
Duke Uniuersrtv, Durham, NC 2 7706, USA Received 9 August 1985

TO

The paper develops a procedure for finding a discrete-valued Markov chain whose sample paths approximate well those of a finance, and econometrics where vector autoregression. The procedure has applications in those areas of economics, approximate solutions to integral equations are required.

1. Introduction
There is interest in economics, finance, and econometrics in the solutions to functional equations where the arguments of the solution functions are the values of an autoregressive process. A typical problem is to characterize the price of an asset, where the law of motion for the dividend is a logarithmic AR(l) process. For example, with an additively separable intertemporal utility function the functional equation for the asset price p is

(1)
where u’(h) is the marginal utility of consumption, h is the dividend, p is the subjective rate of discount, and F( h’ 1h) is the conditional distribution of the dividend. The law of motion for the state variable y = log(h) is y’ = Xy + C, where c is white noise. Under regularity conditions (1) admits a unique solution for the asset price as a function p(v) of the log of the dividend. Such pricing functions are studied by Lucas (1978), Brock (1982) and others. Generally the solutions to (1) are not available in an analytically simple closed form. Instead, the solutions are given as the limit of a sequence of computationally difficult iterations motivated by contraction mapping theorems. However, there are instances where calculation of the exact solution or a good approximation to it - is important. For example, Prescott and Mehra (1985) examine the quantitative aspects of asset pricing in their study of the puzzle of ‘high’ equity returns, while Donaldson and Mehra (1984) study the qualitative features of multivariate asset pricing functions. The strategy adopted in these papers and in other work is to use a finite-state discrete Markov chain for the state variables and to restrict the number of possible values of the state variables to be very small, usually only two or three. In the discrete case the problem of solving the functional equation (1) becomes the simpler problem of just inverting a matrix. If the range space of state variables is small, then one can find ad hoc, though presumably ‘realistic’ numerical values for the transition probabilities and the state variables. However, the
0165-1765/86/$3.50 0 1986, Elsevier Science Publishers B.V. (North-Holland)

I =F 0.‘/(1 . the inversion of very large matrices is practicable. the method works well in an ongoing project [Tauchen (1985)] that investigates the small sample properties of generalized method of moments estimator [Hansen (1982)]. A method for selecting the values y’ is to let YN be a multiple m of the unconditional standard deviation Us = (a. Thus it is possible to use the method to calculate explicitly the solutions to functional equations like (1) using dynamics for the state variables that are close to those actually observed in the economy. under regularity conditions. AY. This is the Kantorovich approach [Wouk (1979. < j” denote the values that jr may take on. For each j.‘. i _F i yk-hy’-w/2 0. (2) where E. This paper develops a method for choosing values for the state variables and the transition probabilities so that the resulting finite-state Markov chain mimics closely an underlying continuousvalued autoregression. or will soon be available [National Science Foundation (1984)]. the size of the matrix to be inverted will become much larger.~ = Pr[ yk . Then let y’ = -yN.w/2 < XyJ + tr <yl‘ + W/2] =F i otherwise. . < U] = F(u/u.-. Then the assignments (3) for p/k make the distribution of jjr conditional on jj. The scalar case Let Y. evidence is presented indicating that the approximation error from the method should be small for moderate sized state spaces.~ = 1. as the state space becomes finer and finer the solution to the functional equation will. Let jjl denote the discrete-valued process that approximates the continuous-valued process (2) and let y’ <y’. -N . after an adjustment for trend. i? and P.. = j ‘1 follows. be generated Y. yN]. The method for calculating the transition matrix plk = Pr[ jt = jh 1j. =y’ be a discrete approxima- . . . Let the distribution function of cr be Pr[c. Below. = by the autoregressive scheme. set Y P. become arbitrarily close to the solution for the continuous case.F The rationale for this assignment of the transition probabilities can be understood by considering a random variable of the form v = AjjJ + c where XYJ is fixed and c is distributed as c.xy’ + w/2 P. Of course. and let the remaining be equispaced over the interval [y’. The motivation for the method is the well-known fact that the statistical properties of many economic time series are captured adequately by vector autoregressions. where F is a cumulative distribution with unit variance. 2.w/2 Y 0. but with the large-scale computational resources that are. This suggests using finer state spaces.).. y’ ._ .xyi . Tauchen / Mm-km-chain approximations difficulty with using a small state space is that one is never sure that unusual and interesting characteristics of the solution are not simply artifacts of the coarseness of the state space.-I + et.1..178 G. pp. In fact. ! y”-hyJ+ww/2 0. is a white noise process with variance 0. Put w = jh -k-‘. Also. if k is between 2 and N . 120&142)] to solving functional equations.X2))‘12.

are computational speed and numerical stability.09) _F~= 0.03 + 0. Monte Carlo studies showed that fitting standard linear time series models to the discrete-valued approximating process j.291 (1) (2) (3) (4) 9 9 9 5 0.1. The advantages of the above scheme. = XJJ. NOBS= 50.176 0..)r2. Gaussian quadrature. s 2 = 0.84 j(. (44 s2=0.101 0.-. = J’ will approximate the sense of weak convergence that of y. Experimentation showed that the quality of the approximation remains good except when X is very close to unity. (0. e.03 + 0. This would necessarily entail the inversion of very large Vandermonde matrices. given y. The parameters x and uz are functions of the second moments of the j( process.229 of the real line formed by the Jk tion to the distribution of the random variable U. and these moments can be computed from the transition matrix and the { 7’).). may lead to a placement of the grid points that in principle is more efficient than the equispaced scheme outlined above.. Tauchen / Markou-chain approxmations 179 Table 1 Case number Number of grid points N Continuous x process 0. The above scheme.-. then the conditional distribution of jr given j. Table 1 shows the induced population statistics x and yF for a range of values for X and N with of = 0. The transition probabilities were computed under a normality assumption for E.15) (0.10 0.14) where standard errors are in parentheses. on the other hand.229 0.0096.100 0. and oc2= (1 -x2) var(j.i..103 0. for the parameters in case no.08 j.90 0.80 0. however.02) (0. = Z.932 oi 0.167 0. A rule based on Gaussian quadrature would essentially use the method of moments to determine the grid points and the transition probability matrix. If the partitioning closely in is reasonably fine.798 0. gives results very similar to what one would expect if the models had been fitted to realizations of the continuous-valued process y. jr.78 j. where x = COV(~~~. .)= 0. admits a representation of the form.g.. can be coded very easily and the approximating Markov chains have been found to be quickly computable for a large number of sets of parameter values for the underlying autoregression.898 0. To assess the adequacy of the approximation (3) note that the discrete process j.G... jr_l)/var(jt). especially in the multivariate case given below. The approximation of x and oF to h and Us is clearly adequate for most purposes when N = 9. NOBS=49._.253 0. a problem which is notoriously time consuming and numerically unstable. indicating that the distance in the K-S metric between the error distribution and the normal distribution is not large.pl. Generating 51 pseudo observations on . These regressions are about what one would expect to get if the continuous-valued process yt had been simulated. Kolmorgov-Smirnov tests accept normality of the residuals.90 0.-.0095. (2) in table 1 and then fitting linear autoregressive models to these data gives J = 0. j. 0. (0. and using the value m = 3 for determining the grid width. Interestingly. It is recognized that other integration rules. with cov(c. itself. Discrete process x 0.02) (0. -xj.

..... For each i let h. I)=~(~~-~. The remaining J..._. 0. N... the products of the appropriate h’s. where 6 is a standardized distribution function. There are N* = N. .(_$-~..1.W -!%+w.:.for be the M x 1 vector of values for the J’s when To calculate transition probabilities pJk let J(j) the system is in state j.. a diagonal matrix.r and y.ii(k)) forj.’ satisfy J. k) = Pr[in state k ]in state j] are. + 0.+1 + 61~~ y2...332 [ 0...)). Let jj.11 for 1 < I < N. wherep=i.* < .. I= 1.:. . N.(r . were each set to nine. of et are mutually independent with distribution Pr[e.:.126 0. A is an M X M matrix. in (5).(j.p.1 the components J.(j.. ..50~~~1 + ~2~. M./2) =F.(u/a(c. < J. After taking appropriate linear combinations any VAR model can be put in the form (5). Tauchen / Markoo-chain appromnatmts 3. where w. (5) where y. which.20~. .NM possible states for the system.126 0.. Each component J..) are the square roots of the diagonal elements of the matrix ZY that satisfies Z’. 2. Let i(j) be an M x 1 vector of integers associated with state j such that when the system is in state j at time I./2) if if if 2</<N. ( y. It is assumed that the elements E. .Z:. . . analogously to (3) is taken to be h. and E. N* as a label for the states.-1. Using this method Y.)/(N. Enumerate these states using the index j=l..185 (0.. . values: y.Y-. . =y/ 1 state j at t .The values 7: were computed (8) with m = 3 The N.-w.‘+ ’ = J.(j).. = 0. . takes on one of N. /=l.w/2) Given the h’s the transition probabilities p( j.N. which can be found by iterating Z.2 .As in the univariate case.. values J. The unconditional ’ 1 yielding 81 states for j.. = 2ma(y. . and put p = AJ( j).1) random variables.. denote the approximating discrete-valued vector Markov chain for y.(Y) = AZ. . by independence the C’S. with convergence as r + 00 guaranteed so long as (5) is stationary../2)-F. and c2( are iid normal this y.180 G.. :’ + w.-.. =J.E. = AZ. I= N.70~.1). = 0.Ns.2 .. is an M x 1 vector white noise process.( j. I) = Pr[J. process is Z Y = 0..NJ are set to minus and plus a small integer m times the unconditional standard deviation of y.r CT.30~2. p(j. =A.. of (6) = 1 . k=1.. The vector case Suppose Y. The a(~... the vector process is + er3 var(c.-.p. N*.t_.+w. is an M X 1 vector.. 2 .) =Z. Assume that after taking such linear combinations the elements of C. (7) covariance matrix of where e. J..A’ + Z:. are also independent. k)=[filh. a discrete approximation was taken to the process + 0. < u] = F... assume the i=l.1)A’ + ..

Statistical properties of generalized method of moments estimators of structured parameters using financial market data. Econometrica 46. NC).200 1 (9) The elements of Aare very close to the corresponding autoregressive parameters in (7). Journal of Monetary Economics 13. 102991055. As in the univariate case it is possible to check the accuracy of the approximation by finding the induced representation .. A._.‘~. IL).Conclusion This paper has developed a method for finding a discrete Markov chain that approximates in the sense of weak convergence a continuous-valued univariate or multivariate autoregression. Review of Economic Studies 51. 145-162... Chicago..‘.200 1 =_. (0.])(E[~. = ’ ’ 0.. Wouk. Experimentation showed that increasing moderately the number of grid points N. = 0.. 1984.~. and R._.15.05) (0. J.. NOBS= 50.13) s2 = 0. with the expectations computed using the transition matrix. Donaldson. 1985. In this case.07 + 0. Asset prices in a production economy. and R. New York). s2=0. + ?. McCall. Michener.. Journal of Monetary Economics 15. to’ these data gives . Asset prices in an exchange economy. Durham.E. Lucas.08) + 0.F. = 0. often improved the accuracy of Zi. Generalized method of moments estimators. Hansen. Access to supercomputers (NSF Office of Scientific Computtng.]). in (8). 491-508. where A= (E[_~. ed.139 0. Permanent income in general equilibrium.17) + 0.44 j2. R. Jr.‘.. Mehra._. Comparative dynamics of an intertemporal asset prtcing model.11 + 0.. Mehra. George.16 j.499 0. Tauchen. 1978.B. The economics of tnformation and uncertainty (The University of Chicago Press. though those of . DC). Markov chain imitates quite well 4. 1979. 50.. A first course in functional analysis (Wiley. Econometrica 50.. Washington.. 1984.11) . Generating 51 observations on this 3..09.standard deviations.i. 1982. Working paper (Duke University.139 0. References Brock. 142991445. R.61 j.. W. The method should be useful in both economics and finance where discrete state spaces are used for finding numerical solutions to integral equations.~. in: J. National Science Foundation. Prescott.299 0.P12. (0. E.34 A. . 279-305.+. The equity premium: A puzzle._.. 1985..p. 0. 1982. NOBS= Comparing the estimates here with (7)-(9) shows that the discrete the statistical properties underlying a first-order vector process.373 0.06) (0.Ff= xj. (0.. and fitting a VAR. 1984. (0.Z? are not quite as close to those of 2. Lars P.